Your brain doesn’t predict language the way AI does

In the last two posts, we looked at how the brain holds onto experience — how sleep consolidates what we've learned, and how emotional intensity stamps certain moments into memory more deeply than others. But memory isn't only about the past. A great deal of what the brain does with stored knowledge is use it to anticipate what comes next. Nowhere is this more striking than in language.

Every time someone speaks to you, your brain is doing something quietly remarkable. Before a sentence is even finished, it has already begun predicting the words to come — drawing on grammar, context, and a lifetime of language experience to stay a step ahead. It happens so automatically that we rarely notice it. And for years, scientists assumed this made the human brain rather like a language model — the technology behind tools like ChatGPT. Both seem to work by anticipating the next word. It's a compelling parallel. A study published this month in Nature Neuroscience suggests it is also, in an important way, wrong.

The assumption that needed testing

Large language models (LLMs) are trained almost exclusively on one objective: predict the next word, given everything that came before. And because these models have shown surprising alignment with patterns of human brain activity, researchers assumed the two systems shared a common strategy. The brain predicts; the model predicts. Same task, same logic.

But the researchers behind the new study — Jiajie Zou, David Poeppel, and Nai Ding — noticed a question that had never been directly tested: does the brain predict each upcoming word as precisely as possible, at every point, in the way language models do? Or does it do something more structured?

What the research found

To find out, the team used magnetoencephalography (MEG) — a brain scanning method that tracks neural activity with millisecond timing — while Mandarin Chinese speakers listened to natural speech. They also analysed a separate dataset of brain recordings from English speakers, allowing them to check whether any results held across two different languages.

Their approach centred on a simple but revealing question: does the brain predict some words more intensely than others? To test this, they used a large language model to calculate how unexpected each word in a sentence was, given everything that came before it. A word that is hard to anticipate from context produces a stronger brain response than one that is predictable. By mapping that response across the full arc of a sentence, the researchers could see exactly where the brain was working hardest to predict what came next.

The answer was clear — and not what a pure next-word prediction model would suggest. Take a sentence like "The exhausted traveller finally reached the hotel." The brain's predictive effort is not spread evenly across each word. Instead, it intensifies within natural phrase-sized chunks — "the exhausted traveller", "finally reached the hotel" — and partially resets at the boundaries between them. These chunks are what linguists call grammatical constituents: the phrases and clauses that the brain naturally groups together as units of meaning. The brain predicts most actively within these units, and recalibrates as it crosses from one to the next.

The effect was also sensitive to ambiguity. When the boundary between chunks was unclear — when listeners couldn't be sure a phrase had ended — the brain's predictive pattern shifted accordingly. The chunking was not rigid; it tracked the structure of the sentence as it unfolded in real time.

What this means in practice

For understanding how language works in the brain: This finding challenges something widely assumed. Rather than predicting word-by-word with equal effort throughout — the way a language model does — the human brain manages its predictive resources according to grammatical structure. It groups language into chunks first, then predicts most actively within them. As co-author David Poeppel puts it: "With LLMs, predictions are by and large created equally. By contrast, the human brain makes predictions by first taking into account chunks of words."

For AI and language model development: If the brain is genuinely more sensitive to grammatical structure than current models, this matters for how we design and evaluate AI language systems. A model that better reflects the brain's constituent-sensitive strategy might align more closely with how humans actually understand speech — and more clearly flag where the two systems fundamentally diverge.

For clinical neuroscience: Language processing difficulties appear across a range of neurological and psychiatric conditions, including aphasia, schizophrenia, and autism spectrum disorder. A clearer account of how the healthy brain structures its predictive effort could provide a more precise framework for understanding what breaks down in these conditions, and at what level.

The key takeaways

• The human brain does not predict the next word with equal effort at every point in a sentence
• Prediction intensifies within grammatical chunks — natural phrase-sized units of meaning — and partially resets at their boundaries
• This pattern was found consistently across Mandarin and English speakers, using two different brain recording methods
• Large language models lack this sensitivity — they predict word-by-word, regardless of phrase structure
• The brain uses memory of grammatical structure to organise its predictive effort, not just to predict the next word in isolation
• The findings suggest that next-word prediction, while real, is managed by a higher-level linguistic strategy that current AI does not replicate

Why this matters beyond the lab

We are living through an unusual moment in which the question of how closely AI resembles the human brain has moved from academic curiosity to something with real cultural and commercial stakes. The assumption that brains and language models share a core computational strategy has shaped research agendas, public narratives, and a great deal of hype about what AI can and cannot do.

This study doesn't overturn that alignment entirely. The brain does predict upcoming words, and language models do capture something genuine about how that works. But the differences matter. The brain is not running next-word prediction on a raw stream of words — it is parsing that stream into structured chunks and allocating its predictive effort accordingly. That is a meaningfully different strategy, one rooted in the same capacity for chunking and compression that makes memory possible in the first place.

The brain doesn't just predict the next word. It understands. Whether AI does the same is a question this fields is only beginning to ask honestly.

This is a lay summary of research published in Zou et al. (2026), drawing on neuroimaging studies of language prediction in Mandarin Chinese and English speakers. For the primary study, see: Zou, J., Poeppel, D. & Ding, N. (2026). Constituent-constrained word prediction during language comprehension. Nature Neuroscience. https://doi.org/10.1038/s41593-026-02272-6. For relevant background, see: Goldstein, A. et al. (2022). Shared computational principles for language processing in humans and deep language models. Nature Neuroscience, 25, 369–380. https://doi.org/10.1038/s41593-022-01026-4